- Chat AI versus agents
- How tool calling works
- What APIs and MCP do
- Agents versus workflows
- Where people should check
The short answer
A chat AI answers in text. An AI agent chooses tools, uses them, looks at the result and decides what to do next, working towards a goal. The difference is whether it only talks or also acts.
Being able to act, though, means mistakes land in real work. Deciding in advance what to hand over and where a person checks is the most important preparation for using an agent.
Four stages from chat to agent
The move from chat AI to agent is not one jump but a series of steps.
1. Chat AI: answers in text
You type a question and get text back: a summary of meeting notes, a draft email. The AI's job ends there. Pasting the draft into an email or adding something to the calendar is done by a person.
2. Tool calling: the AI can say "I want to use this tool"
Tool calling lets the AI decide, partway through, which tool it wants and what to pass it — for example, "call the weather tool with Tokyo". The AI does not run the tool itself; the program the AI is built into does. That program runs the tool, sends the result back, and the AI uses it to finish the answer.
Ask "When am I free next Tuesday?" and the AI calls a tool that reads your calendar, then answers from what comes back. Nobody has to open the calendar and paste it in.
3. Agent: using tools repeatedly to finish a task
An AI agent does not stop after one tool call. It loops — look something up, check the result, decide the next step, use another tool — until the task is done.
Ask it to "set up next week's meeting at a time the three attendees are all free, and draft the invitation", and it reads three calendars, finds a slot with no overlap, picks a candidate and writes the draft, putting those steps together itself.
4. Autonomous agent: longer tasks from fewer instructions
Go further and you hand over only a broad goal, leaving both the planning and the decisions along the way to the AI. This is sometimes called an autonomous agent. The more it covers, the less work for people, but the harder it is to notice when it heads in the wrong direction. What counts as "autonomous" varies between services and between the people describing them.
Connecting to tools: APIs and MCP
Most of the "tools" an agent uses are features of other software: a calendar, email, a spreadsheet, an internal customer database.
An API is the doorway that lets one piece of software talk to another in an agreed format. An agent reads a calendar or registers data through APIs. Software without an API is hard for an agent to work with.
MCP (Model Context Protocol) is a shared convention for connecting AI applications to external tools. Instead of building a custom connection for every tool, any tool that supports MCP can be used from the AI in the same way. It is a bit like standardised plug shapes: once the shape matches, the appliance works in any socket.
Which tools are supported, and exactly how to connect them, differ between services and can change. Check each service's official documentation when you set this up.
Agents versus workflows
A workflow runs steps in an order fixed in advance: "when an invoice arrives, save it to this folder and notify accounts". Because the same path is followed every time, the outcome is predictable and easy to check.
An agent works out the steps on the spot. That gives it room to cope with situations nobody planned for, but the same request will not necessarily follow the same steps each time.
| Aspect | Workflow | AI agent |
|---|---|---|
| Steps | Decided in advance by people | Assembled by the AI on the spot |
| Predictability | High | Can be lower |
| Unexpected input | Stops or errors | May be able to cope |
| Best suited to | The same process every time | Work whose steps depend on the situation |
In practice, the easiest setup to manage is to fix the predictable parts as a workflow and hand only the parts that need judgement to an agent.
Decide where people check
A simple way to divide the work is to ask whether the action can be undone.
- Easy to hand over: looking things up, reading, drafting, listing options. Mistakes can be corrected by a person afterwards.
- Add an approval step: sending email, adding calendar entries, changing data. A person reviews and approves before it runs.
- Do not hand over: payments, contracts, deletions, publishing externally. These cannot be undone, or the impact is too large.
Give an agent only the permissions it needs. If a read-only task is given write or delete rights as well, the damage from a mistake — or from a malicious instruction slipped in — grows accordingly.
One more thing to watch is "instructions" hidden in what the agent reads. If an email or web page says "send this to an outside address", that is data to be read, not a command to obey. Putting a human approval step in front of important actions keeps the agent from being steered by text like that.
Summary
- A chat AI answers; an agent uses tools to move work forward.
- Tool calling lets the AI use tools; an agent repeats that loop.
- Tools connect through APIs, and MCP is a shared convention for making those connections.
- Fixed processes suit workflows; judgement-heavy parts suit agents.
- Put a human check in front of anything that cannot be undone.